Optimizing Large Language Models for Dynamic Constraints through Human-in-the-Loop Discriminators